Z27AG decode pass 3: wheel positions from corners, odometer (570, 24-bit); layouts can be 24 bits

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDwfd6Ki8UfG4hK6W5YpJz
2026-10-15 22:13:34 +09:00
parent 642be2a5e1
commit e8ba581985
3 changed files with 157 additions and 3 deletions
@@ -17,7 +17,7 @@ ground truth. Prior art and first candidates: [the prior-art report](../reports/
 | `vehicle.ndjson` | `decoder` and `obd-poller` via the bus | decoded CAN signals, and the `obd_*` references from the K-line through the ELM327 (rpm, speed, throttle, load, MAF, timing, O2 every round; coolant, intake air, trims every 5 s); `env.timestamp` is the reply's arrival at the Orin |
 | `events.ndjson` | marks from the app | when each step happened |
 | `gnss.ndjson` | gpsd (SIM7600) | speed and position for the drive |
-| `imu.ndjson` | `imu` via the bus (`yukart/raw/imu`) | every IMU sample, raw module axes, Orin receive time: calibration (27–30) and, once calibrated, motion for the drive |
+| `imu.bin` (`imu.ndjson`) | `imu` via the bus (`yukart/raw/imu`) | every IMU sample, packed (svc-core `imu --export` prints it as JSON lines; sessions before 2026-10-16 and anything that does not pack are in `imu.ndjson`), raw module axes, Orin receive time: calibration (27–30) and, once calibrated, motion for the drive |

 ## Operating it from the app

@@ -64,7 +64,7 @@ recorder refuses marks with "no session" otherwise); engine off, doors shut.
  twice on the same spot, facing one way and then turned round (a rough 180° is fine:
  the ground's slope reverses and cancels, the mounting's tilt does not); then three
  firm pull-aways and three firm stops on a straight, flat section (forward, signed by
-  which kind of stretch it is). Any session holds raw IMU samples (`imu.ndjson`), so
+  which kind of stretch it is). Any session holds raw IMU samples (`imu.bin`), so
  these can go in their own short session or with anything else.

 Version 3 (2026-10-14) adds the IMU calibration section. Version 2 (2026-10-13) adds the stands section and writes down the conditions the
@@ -0,0 +1,59 @@
+---
+type: report
+date: 2026-10-15
+---
+# Z27AG CAN: third pass, the drive back
+
+Sessions `000011-20261015T212904` and `000013-20261015T214544` (the drive back, about 13 minutes,
+no marks) together with pass 2's `000008` and `000009`: 20 km of driving with GNSS and the calibrated
+IMU, no OBD-II. Nothing was marked, so this pass uses what any drive has: corners and distance.
+
+## Confirmed
+
+| Signal | Where | Evidence | Registered as |
+|---|---|---|---|
+| Wheel speeds | `235` four words, little-endian (pass 2): bytes 0–1 **front right**, 2–3 **front left**, 4–5 **rear right**, 6–7 **rear left** | from corners, below | `wheel_speed_fr_kph`, `_fl_`, `_rr_`, `_rl_` |
+| Odometer | `570` bytes 4–6, little-endian, **1 km** | steps once per kilometre: 12 steps over 11.9 km (`000009`), 3 over 3.0 km (`000011`); 167966 on the 13th, 167988 at the end of the drive back; valid with the engine off too | `odometer_km` (the layout format gained 24-bit fields) |
+| Distance counter | `14D` bytes 6–7, little-endian, **about 0.40 m** per count | tracks distance from the wheel speeds with correlation 1.0000 in both drives (103.2 and 103.6 m per step of byte 7); wraps every ~26 km | not registered: the odometer covers whole kilometres; this is there if a trip meter needs finer steps |
+
+### Which word is which wheel
+
+In a corner the outer wheels run faster than the inner ones, by about half the track width times the
+curvature; and the fronts run on a wider arc than the rears, by about half the wheelbase squared times
+the curvature squared. Fitting each word's share of the four (its speed ÷ their mean) against the
+curvature from the steering angle (`steering_angle_deg` + 2.7° for the straight-ahead offset, ÷ 16.3,
+over a 2.5 m wheelbase), 74 476 frames above 15 km/h:
+
+| Word | Linear term (track) | Square term (wheelbase) | So |
+|---|---|---|---|
+| 0 | +0.72 m | +1.4 m² | right, front |
+| 1 | −0.72 m | +1.4 m² | left, front |
+| 2 | +0.75 m | −0.9 m² | right, rear |
+| 3 | −0.75 m | −1.9 m² | left, rear |
+
+Curvature is positive turning left, so a positive linear term is a wheel on the right (outside in a left
+turn). The size, ±0.72–0.75 m, is half the Colt's track (1.45–1.46 m). The square terms split the words
+into a pair above the mean and a pair below it, ±1.5 m² against an expected ±1.6 m² (2.5² ÷ 4). Medians
+by curvature agree: in left turns tighter than 20 m (`κ > 0.05`) the words read 1.068, 0.956, 1.044,
+0.932 of their mean.
+
+Straight ahead the fronts read 0.6 % more counts than the rears, matching pass 2's per-word scales
+(0.0578 against 0.0581 km/h per count). Driven wheels or tyre sizes, either is ordinary; each pair is
+registered with its own scale, so all four read GNSS speed when cruising.
+
+Jack stands (checklist 23–26) would still confirm this, but are no longer needed to find it.
+
+## Candidates
+
+- **`570` bytes 2 and 3** move together (byte 3 ≈ byte 2 + 48) and fall slowly: byte 2 read 87 on the
+  13th and 76 at the end of the drive back, 22 km later. Fuel level fits the fall, but byte 2 also rose
+  by about 3 while parked between drives (78 → 81), which fuel would not do on level ground. It may
+  be outside temperature instead, falling through the evening. Both read `7F` with the engine off. A
+  refuel settles it: fuel jumps, temperature does not. Not registered.
+
+## Notes on the recording
+
+- `000012` is a 25-second parked session in between; `000013` ends with the PEAK unplugged.
+- The recorder kept IMU samples as JSON lines (34 MB for `000009`). From this pass on it packs them
+  (`imu.bin`, svc-core `imu/packed.rs`): the same samples are 8.5 MB, and all of `000009`'s samples fit
+  the packed form exactly.